In this section, we discuss the possible energy savings for each of the three buildings by using the knowledge of the heat load forecast provided by our model. From the results of the heat load forecast we obtain the best case forecasting accuracy for HL(t+1). Therefore, we estimate the energy savings for the buildings for the forecast horizon of one hour HL(t+1) to illustrate the benefits of using a forecasting model for energy savings.
To estimate the energy savings for the buildings we make a couple of assumptions. We assume that each building has been assigned a maximum heat load limit per hour,HLmax considering the physical features and number of occupants residing in the building. We also assume that in absence of the knowledge of heat load forecast the energy producer needs to supply this maximum heat loadHLmax for the building for each hour to satisfy the heat energy demand of the building.
The maximum heat load consumption for each building during both winter and spring season can be observed from Table 6 and Table 7. We consider Building A during winter season to compute the energy savings. We observe that maximum heat load limit per hour HLmax is 124,533 Watts. The model achieves a forecasting accuracy of 82.29 % forHL(t+1)for Building
A during winter season using EWD and Naive Bayes. The discrete states of heat load are repre-sented in Table 9. The number of correct and incorrect predictions for each state is reprerepre-sented by the confusion matrix shown in Table 11. State A is correctly predicted 446 times, State B 429 times and so on. When the model predicts a particular state for the next hour, the heat energy needed to be supplied by the energy company is the maximum value of the particular state so as to satisfy the heating demand for that particular hour. We compute the energy savings in case of correct predictions by the model for each state. This is achieved by subtracting the maximum heat load of a particular state from the maximum heat load limitHLmaxof the Building A. This difference is multiplied by the number of times that particular state is correctly predicted by the model(number of correct predictions of each state also represents the number of hours during which that state was predicted)
Total energy savings for Building A during winter season
= 446∗(124533−48410) + 429∗(124533−67436) + 239∗(124533−86471)+
192∗(124533−105235) + 56(124533−124533)
= 71,247,505 W atts
Heat energy spent in absence of heat load forecast information = Sum of all elements of confu-sion matrix(for both correct and incorrect predictions) *HLmax
= 1655∗HLmax
= 1655∗124,533 = 206,102,115 W atts
Therefore, percentage of energy savings due to correct forecasts = 71247505/206,102,115 ∗ 100 = 34.56%
Thus, with the correct prediction of heat load forecast energy savings of around 35 % can be obtained in Building A during the winter season. However, the accuracy of the forecast model was 82.29 %. In cases of incorrect predictions of state of heat load forecast, the proposed model both overestimates and underestimates the heat load consumption in buildings. Due to the presence of both overestimation and underestimation scenarios with several instances, it is computationally intensive to consider all these cases and compute their impact on the energy savings. However, the energy savings obtained from the correct number of predictions highlight the fact that our model leads to significant energy savings in buildings. The table below shows the percentage of energy savings obtained from the correct forecasts in all three buildings across both seasons.
Table 17.Energy savings forHL(t+1)forecast for correct predictions
In the previous section, we computed the energy savings for each building by using the knowl-edge provided by the heat load forecast. The heat load forecast tells the production company the amount of heat energy it needs to produce for a particular period of time. Thus, the energy production company does not need to produce excess heat energy which makes them energy efficient and leads to energy savings. It also lowers the consumption of fuel needed to produce heat energy. This leads to efficient utilization of natural resources. The idea of sustainable development is to use the available resources efficiently to meet the needs of the present genera-tion without compromising the needs of the future generagenera-tion. Therefore, the heat load forecast information provided by our model helps in achieving sustainability. By using the heat load fore-cast information buildings will consume less energy and this will also decrease the greenhouse gas emissions. Further, energy savings also lead to financial savings for the energy production company. This ultimately also leads to reduced costs for the building occupants. These energy efficient buildings thus contribute to sustainable and energy efficient cities.
4.5 Summary
In this chapter, we discussed the results of the heat load forecast for the proposed model. We evaluated and analysed the performance of the model over two seasons and different horizons in three residential buildings. By computing the average accuracies across both seasons we observed that current heat load consumption and outdoor temperature forecast are the two pa-rameters with the most influence on the heat load forecast. We also investigated the classification errors and explained the reason behind the resulting accuracy of the model. We also observed that model achieves a good accuracy for forecasting the heat load for the next hour by using less training data. We estimated the energy savings using the heat load forecast information for HL(t+1)in all three buildings. In the next chapter, we conclude the thesis work and discuss the future research directions.
5 CONCLUSION AND FUTURE WORK
In this chapter, we discuss the conclusion and future work related to this thesis work. The limi-tations of the Bayesian approach are also discussed. The objective of the thesis was to develop heat load forecasting models using the Bayesian approach. The heat load forecast in buildings was computed by studying the impact of several parameters on the heat load consumption by developing a machine learning model based on a Bayesian network.
5.1 Conclusion
This thesis work presented a Bayesian approach for forecasting the heat load in residential build-ings in a district heating system. We studied that forecasting the heat load consumption helps in optimizing the heating production. The district heating operation was studied to identify the parameters influencing the heat load consumption in residential buildings. The parameters iden-tified for heat load forecast included DHS operational parameters (supply temperature, return temperature, flow rate, difference between supply and return temperature), outdoor temperature forecast, behavioural parameters (hour of day and day of week) and current heat load consump-tion. We considered the Bayesian inference methodology due to its advantage of assessing the probability of uncertain and non-repeateable events. We modelled a Naive Bayes network and used two discretization techniques for converting continuous attributes to discrete states, in our proposed model. The forecast model was built by utilizing the realistic district heating data over a period of 4 months across winter and spring seasons from three residential buildings in Skellefteå, Sweden. Heat load forecasting was performed for horizons of 1, 2, 3, 6 and 24 hours to consider the effect of the district heating control loop and daily heat load consumption pattern.
Our results indicate that the current heat load consumption and outdoor temperature forecast are the two most important parameters influencing the heat load forecast. In this case, our model achieves average accuracies of 81.23% and 76.74% for a forecast horizon of 1 hour, HL(t+1) in three buildings for winter and spring seasons, respectively. We also observe that the combined influence of DHS operational parameters and outdoor temperature forecast on the heat load forecast is more significant than their individual influence. Further, by utilizing only 10% of training data, our model was able to achieve an average accuracy of 77.97% for the three buildings across both seasons with forecast horizon of 1 hour,HL(t+1). We observed that the forecasting accuracy in Building A was higher than in Buildings B and C because of less variation of heat load in Building A. We analysed the results of the forecasting accuracy
by diagnosing the probabilities of various states of heat load forecast by setting evidence to particular states of outdoor temperature. We concluded that classifier errors are mostly due to the high variation in heat load consumption for the same values of outdoor temperature. We also conclude that by utilizing the Bayesian approach we were able to forecast the heat load for the next hour with a good accuracy by just utilizing the 10-20 percent of the training data. We believe from our results, that in case of a large number of buildings and large amount of data, our model would be suitable for heat load forecast. The estimated energy savings in all three buildings show that the heat load forecast information predicted by the model plays a key role in making buildings sustainable and energy efficient.